Events
Take the Microsoft Learn Challenge
Nov 19, 11 PM - Jan 10, 11 PM
Ignite Edition - Build skills in Microsoft Azure and earn a digital badge by January 10!
Register nowThis browser is no longer supported.
Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.
APPLIES TO: Azure CLI ml extension v2 (current) Python SDK azure-ai-ml v2 (current)
This article describes how Azure Machine Learning uses machine learning operations (MLOps) to manage the lifecycle of your models. Applying MLOps practices can improve the quality and consistency of your machine learning solutions.
MLOps is based on DevOps principles and practices that increase the efficiency of workflows, such as continuous integration, continuous deployment, and continuous delivery. Applying these principles to the machine learning lifecycle results in:
MLOps provides the following capabilities to the machine learning process:
For more information on MLOps, see Machine learning operations.
Use Azure Machine Learning pipelines to stitch together all the steps in your model training process. Machine learning pipeline steps can include data preparation, feature extraction, hyperparameter tuning, and model evaluation.
In the Azure Machine Learning studio Designer, you can clone a pipeline to iterate over its design without losing your old versions. To clone a pipeline at any time in the Designer, select Clone in the top menu bar.
For more information on Azure Machine Learning pipelines, see Machine learning pipelines.
Azure Machine Learning environments ensure that builds are reproducible without using manual software configurations. Environments can track and reproduce the pip and conda software dependencies for your projects.
You can use environments for model training and deployment. For more information on environments, see Azure Machine Learning environments.
Azure Machine Learning can use MLOps from anywhere to register, package, and deploy models.
Model registration stores and versions your models in your Azure Machine Learning workspace in the Azure cloud. The model registry makes it easy to organize and keep track of your trained models.
A registered model is a logical container for one or more files that make up your model. For example, if your model is stored in multiple files, you can register the files as a single model in your Azure Machine Learning workspace. After registration, you can download or deploy the registered model and receive all the component files.
You can also register models that are trained outside of Azure Machine Learning. Azure Machine Learning supports any model that can be loaded by using Python 3.5.2 or higher.
You identify registered models by name and version. Whenever you register a model with the same name as an existing model, the registry increments the version number.
You can provide metadata tags during registration and use these tags to search for a model.
Important
You can't delete a registered model that's being used in an active deployment.
For more information on how to use models in Azure Machine Learning, see Work with models in Azure Machine Learning.
To deploy a model into production, you must first package it into a Docker image. In most cases, image creation automatically happens in the background during deployment. However, you can manually specify the image.
It's useful to deploy to your local development environment first so you can troubleshoot and debug before deploying to the cloud. This practice can help you avoid having problems with your deployment to Azure Machine Learning. For more information on how to resolve common deployment issues, see How to troubleshoot online endpoints.
You can convert your model to Open Neural Network Exchange (ONNX) to try to improve performance. Typically, converting to ONNX can double performance.
For more information on ONNX with Azure Machine Learning, see Create and accelerate machine learning models.
You can deploy trained machine learning models as endpoints locally or in the cloud. Deployments use CPUs and GPUs for inferencing.
To deploy a model as an endpoint, you need to provide the following information:
Important
When you deploy an MLflow model, you don't need to provide an entry script or an environment for the deployment. For more information on deploying MLflow models, see Guidelines for deploying MLflow models.
You also provide the configuration of the target deployment platform, such as the virtual machine (VM) family type, available memory, and number of cores. When Azure Machine Learning creates the image, it also adds any components it needs, such as assets needed to run the web service.
Batch scoring is supported through batch endpoints. For more information on batch scoring, see Batch endpoints.
You can use your models with online endpoints for real-time scoring. Compute targets for online endpoints can be local development environments, managed online endpoints, or Azure Kubernetes Service (AKS).
To deploy a model to an online endpoint, you need to provide the following information:
For more information on deployment for real-time scoring, see Deploy online endpoints.
When you deploy to an online endpoint, you can use controlled rollout to enable the following scenarios:
For more information on deployment using a controlled rollout, see Perform safe rollout of new deployments for real-time inference.
Azure Machine Learning gives you the capability to track the end-to-end audit trail of all your machine learning assets by using metadata. For example:
Some information on models and data assets is automatically captured, but you can add more information by using tags. When you look for registered models and data assets in your workspace, you can use tags as filters.
Note
When you use the Tags in the Filter by option on the Models page of Azure Machine Learning studio, be sure to use TagName=TagValue
without spaces rather than TagName : TagValue
.
Azure Machine Learning publishes key events to Azure Event Grid, which can be used to notify and automate on events in the machine learning lifecycle. For more information on how to set up event-driven processes based on Azure Machine Learning events, see Custom CI/CD and event-driven workflows.
You can use Git and Azure Pipelines to create a continuous integration process that trains a machine learning model. In a typical scenario, when a data scientist checks a change into a project's Git repository, Azure Pipelines starts the training job.
You can inspect the job results to see the performance characteristics of the trained model. You can also create a pipeline that deploys the model as a web service.
The Machine Learning extension makes it easier to work with Azure Pipelines. The extension provides the following enhancements to Azure Pipelines:
For more information on using Azure Pipelines with Azure Machine Learning, see Use Azure Pipelines with Azure Machine Learning.
Microsoft Power BI supports using machine learning models for data analytics. For more information, see AI with dataflows.
Events
Take the Microsoft Learn Challenge
Nov 19, 11 PM - Jan 10, 11 PM
Ignite Edition - Build skills in Microsoft Azure and earn a digital badge by January 10!
Register nowTraining
Learning path
Introduction to machine learning operations (MLOps) - Training
Introduction to machine learning operations (MLOps)
Certification
Microsoft Certified: Azure Data Scientist Associate - Certifications
Manage data ingestion and preparation, model training and deployment, and machine learning solution monitoring with Python, Azure Machine Learning and MLflow.